IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

WA for Benchmark App: Remote tensor use for internaly dynamic model (#12858)

Browse Source 
* WA for Benchmark App: Remote tensor use for internaly dynamic original models (NMS) that turn static due to transformation to legacy op and lack of dynamism support

* Style
This commit is contained in:
Evgenya Stepyreva 2022-09-02 21:14:33 +04:00 committed by GitHub
parent 0ad91f040c
commit ac6e180bdd
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 30 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

36
samples/cpp/benchmark_app/remote_tensors_filling.cpp
View File

@ -128,18 +128,43 @@ std::map<std::string, ov::TensorVector> get_remote_input_tensors(
#endif
}
ov::Shape get_static_shape(const ov::Output<const ov::Node>& compiled_output) {
// FIXME: this is a WA for case when original model has internal dynamism (NonMaxSuppression)
// and runtime has static output due to conversions to legacy op and lack of dynamism support
// to be removed along with dynamism support
const auto& compiled_pshape = compiled_output.get_partial_shape();
if (compiled_pshape.is_static())
return compiled_pshape.to_shape();
else if (compiled_pshape.rank().is_dynamic())
OPENVINO_UNREACHABLE(
"Benchmark App - NOT IMPLEMENTED - Output of dynamic rank is not supported for remote tensor. ",
"Output: ",
compiled_output);
ov::Shape shape;
for (const auto& dimension : compiled_pshape) {
if (dimension.get_interval().has_upper_bound())
shape.push_back(static_cast<ov::Shape::value_type>(dimension.get_max_length()));
else
OPENVINO_UNREACHABLE("Benchmark App - NOT IMPLEMENTED - Fully dynamic output dimensions are not supported "
"for remote tensor. ",
"Output: ",
compiled_output);
}
return shape;
}
std::map<std::string, ov::Tensor> get_remote_output_tensors(const ov::CompiledModel& compiledModel,
std::map<std::string, ::gpu::BufferType>& clBuffer) {
#ifdef HAVE_DEVICE_MEM_SUPPORT
std::map<std::string, ov::Tensor> outputTensors;
std::shared_ptr<const ov::Model> runtime_model = nullptr;
for (auto& output : compiledModel.outputs()) {
auto context = compiledModel.get_context();
auto& oclContext = static_cast<ov::intel_gpu::ocl::ClContext&>(context);
auto oclInstance = std::make_shared<OpenCL>(oclContext.get());
ov::Shape shape = get_static_shape(output);
cl_int err;
auto elementsNum =
std::accumulate(begin(output.get_shape()), end(output.get_shape()), 1, std::multiplies<size_t>());
auto elementsNum = shape_size(shape);
auto inputSize = elementsNum * output.get_element_type().bitwidth() / 8;
cl::size_type bufferSize = 0;
@ -153,9 +178,8 @@ std::map<std::string, ov::Tensor> get_remote_output_tensors(const ov::CompiledMo
buff = cl::Buffer(oclInstance->_context, CL_MEM_READ_WRITE, (cl::size_type)inputSize, NULL, &err);
}
}
outputTensors[output.get_any_name()] = oclContext.create_tensor(output.get_element_type(),
output.get_shape(),
clBuffer[output.get_any_name()].get());
outputTensors[output.get_any_name()] =
oclContext.create_tensor(output.get_element_type(), shape, clBuffer[output.get_any_name()].get());
}
return outputTensors;